Understanding Internet Performance from the User Perspective Project

Summary

To develop measurement and metrics to quantify performance of the Internet from the Customer viewpoint, NIST/ITL joined Cross Industry Working Team (XIWT), consisting of the leading U.S. computer and communication companies. The initial driver was "service level agreements." NIST became a monitoring site to collect data on "round trip delays and packet loss" of Unix Ping Packets exchanged among the participating organizations. The monitoring and data collection role is subsequently transferred to CNRI (Corporation for National Research Initiatives), the administrative organization for XIWT. Since behavior of the Internet is statistical in nature, SED started collaborating with XIWT jointly with ITL Advanced Network Technologies Division. ANTD and SED jointly made a successful proposal to DARPA. The project began with the DARPA funding starting in March, 2000.

Description

As part of a DARPA sponsored project, the SED work is one of the three elements. The SED element involves providing Web based capabilities for exploratory data analysis, doing research on statistical modeling, and contributing to collaborative research among XIWT member companies. This is a very dynamic and fast moving business. The priorities in XIWT are evolving with the interests of the participants. We see XIWT as the driver and SED as supporting the industry initiatives.

Major Accomplishments

ITL has performed the following work for this project.

• Did exploratory data analysis of four months (5/1/98 - 9/4/98) of data from NIST to SLAC-Stanford, Westgroup and Intel. The data consists of source, destination, time packet size, fraction of packet loss, and summary statistics (average, minimum, and maximum) on round trip delay. The exploratory data analysis (EDA) indicated effect of packet size, certain anomalies, weekly trend, correlations. A presentation to XIWT members resulted in synchronizing of the clocks and interest in bi-directional delays.
• Using time series analysis methodologies, we have analyzed the ping delay time data for different destinations and with different ping sizes. We found that the average delay time vary with destinations as well as the size of the ping. For given destination and ping size, extensive data analysis shows that data in general has some dynamic structure. That means that data can be modeled as time series, for better effect the model needs to be updated or modified. We found that data in general has stationary structure with long memory: this means that random fluctuation of the data have some kind of long-term pattern. Further, data also shows persistent periodic behavior with a daily period for most cases. A very important characteristic of the data is that the (marginal) probability distribution of the delay time is not bell shaped or Gaussian. With this distinctive feature in mind, monitoring and predicting the delay time will definitely need tools not based on the ordinary Gaussian assumption.

The collaborators are XIWT member companies. Work is coordinated by CNRI. They have performed the following work for this project.

• The XIWT has written two white papers: Customer View of Internet Service Performance: Measurement Methodology and Metrics; and Internet Service Performance: Data Analysis and Visualization. SED had suggested use of Box plots for the plots used in the second report. That is what was done.
• The XIWT companies are continuing to monitor using exchange of ping packets, expand the network, and improve it. The ping data are sent by the XIWT monitoring sites to CNRI every night. Plans are to duplicate that data at NIST to be integrated with a Statserver, purchased from Mathsoft Inc.